SISA: Set-Centric Instruction Set Architecture for Graph Mining on Processing-in-Memory Systems

Abstract: Simple graph algorithms such as PageRank have been the target of numerous hardware accelerators. Yet, there also exist much more complex graph mining algorithms for problems such as clustering or maximal clique listing. These algorithms are memory-bound and thus could be accelerated by hardware techniques such as Processing-in-Memory (PIM). However, they also come with nonstraightforward parallelism and complicated memory access patterns. In this work, we address this problem with a simple yet surprisingly powerful observation: operations on sets of vertices, such as intersection or union, form a large part of many complex graph mining algorithms, and can offer rich and simple parallelism at multiple levels. This observation drives our cross-layer design, in which we (1) expose set operations using a novel programming paradigm, (2) express and execute these operations efficiently with carefully designed set-centric ISA extensions called SISA, and (3) use PIM to accelerate SISA instructions. The key design idea is to alleviate the bandwidth needs of SISA instructions by mapping set operations to two types of PIM: in-DRAM bulk bitwise computing for bitvectors representing high-degree vertices, and near-memory logic layers for integer arrays representing low-degree vertices. Set-centric SISA-enhanced algorithms are efficient and outperform hand-tuned baselines, offering more than 10x speedup over the established Bron-Kerbosch algorithm for listing maximal cliques. We deliver more than 10 SISA set-centric algorithm formulations, illustrating SISA's wide applicability.

• The paper introduces SISA, a set-centric instruction set architecture that leverages processing-in-memory systems to accelerate complex, memory-bound graph mining algorithms.
• SISA utilizes set operations accelerated by PIM architectures, achieving significant speedups, including over tenfold compared to the Bron-Kerbosch algorithm for maximal cliques.
• This set-centric PIM approach provides a scalable template for improving existing graph analytics frameworks and suggests new ways to design algorithms for future data-centric hardware.
• meta_description
• The SISA paper introduces a set-centric instruction set architecture leveraging processing-in-memory systems to significantly accelerate complex graph mining tasks.
• title
• SISA: Set-Centric ISA for Graph Mining on PIM Systems

Overview of the SISA Paper

The paper "SISA: Set-Centric Instruction Set Architecture for Graph Mining on Processing-in-Memory Systems" presents a comprehensive approach to bridge the performance gap in complex graph mining tasks. Typical graph algorithms such as PageRank benefit from various hardware accelerators, but numerous memory-bound graph mining algorithms, like clustering or maximal clique listing, have not been effectively harnessed by traditional architectures due to their intricate memory access patterns and parallelism. The authors introduce SISA—a set-centric instruction set architecture—to address these challenges by leveraging simple yet efficient set operations that can be mapped to PIM technologies.

Key Contributions

1. Set-Centric Programming Paradigm: The paper introduces a programming paradigm focused on set operations, which allows researchers to express graph mining algorithms as operations over sets of vertices or edges. This abstraction helps in decomposing complex graph algorithms into simpler operations, improving both programmability and execution efficiency.
2. Set-Centric ISA Extensions: SISA provides ISA extensions optimized for set operations such as intersection, union, difference, and membership. This design emphasizes both expressivity and efficiency, enabling graph mining algorithms to leverage processing-in-memory capabilities seamlessly.
3. PIM Acceleration: The paper utilizes two types of PIM architectures to accelerate set operations—bulk bitwise operations in DRAM for densely connected vertices and logic layers in near-memory systems for sparse arrays. This dual approach allows SISA to alleviate bandwidth bottlenecks effectively.
4. Algorithmic Efficiency: SISA-enhanced algorithms demonstrate substantial improvements, with more than ten formulations presented in the paper. Notably, these optimized algorithms outperform hand-tuned baselines, such as achieving over a tenfold speedup compared to the Bron-Kerbosch algorithm for maximal cliques.
5. Cross-Layer Integration: SISA is designed for integration into systems using the RISC-V ISA, ensuring modularity and future adaptability. It offers theoretical work guarantees, making it competitive against traditional graph processing paradigms while providing a new method of expressing and executing graph mining algorithms.

Numerical Results and Claims

The paper reports significant numerical improvements when applying SISA across various graph mining problems. Empirical results showcase that SISA-enabled algorithms consistently outperform existing solutions, proving its wide applicability and potential for broad adoption in graph analytics domains. The integration of PIM enables set operations to harness internal DRAM bandwidth, presenting notable speedups over more than 10 competitive algorithms. The paper supports its assertions through detailed evaluations on diverse datasets, confirming the substantial impact of set-centric programmability combined with PIM.

Implications and Future Directions

The implications of this research are both practical and theoretical. Practically, the approach offers a scalable template for enhancing performance in existing graph analytics frameworks. Theoretically, it opens avenues for rethinking graph algorithm design to prioritize data structure abstraction, potentially influencing subsequent developments in specialized data-centric architectures.

Looking forward, SISA could expand its applicability beyond static graph processing to incorporate dynamic graph analytics, support more complex graph algorithms, and integrate additional hardware accelerators like GPUs or FPGAs to broaden performance advantages. As graph data size and complexity continue to surge, leveraging PIM within a set-centric framework promises significant advancements in handling real-world graph workloads.

In conclusion, the paper sets a foundational precedent for leveraging set-centric paradigms alongside sophisticated ISA designs to boost graph mining performance systematically. By harnessing PIM and effective set operations, SISA offers both theoretical and empirical benefits, marking a significant step in the evolution of hardware-software co-design for graph analytics.